Skip to content

t1k:marketing:ua:campaign-optimization-training

FieldValue
Moduleua
Version1.13.11
Effortmedium
Tools—
/t1k:marketing:ua:campaign-optimization-training
[topic] — e.g. 'the metric', 'ranking', 'cutting', 'scaling', or omit for the full curriculum

Teaches the reasoning behind the studio’s UA optimization decisions. Every other UA skill executes the loop; this one explains why it is shaped that way, so a new buyer can make the call when the situation is not in the runbook.

Scope. Cross-network fundamentals only — the reasoning that transfers to any ad network. Network-specific mechanics (Google policy triage, Mintegral bid levers) live in the execution skills listed under Execution handoff. Monetization, creative production, and attribution setup are out of scope.

This skill decides nothing. It does not read live data or recommend a cut. When the learner wants to act on today’s numbers, hand off — do not improvise a recommendation from the teaching material.

Learner intentPath
“Teach me the whole thing”Run all six lessons in order → Curriculum
“Why 120%? Why this metric?”Lessons 1-2 → Curriculum
“How do I pick what to fix first?”Lesson 3 (ranking)
“How much do I cut?”Lesson 4 (cutting)
“When do I spend more?”Lesson 5 (scaling)
“I want to optimize today”Not training → Execution handoff

Work one lesson at a time. For each: state the rule, give the number that makes it concrete, then ask the learner to apply it to a case before moving on. Do not present all six at once — the failure mode is a learner who can recite the rules and still ranks by percentage.

Worked cases with real-shaped numbers: references/worked-examples.md.

Lesson 1 — The metric is calendar, not cohort

Section titled “Lesson 1 — The metric is calendar, not cohort”

Daily optimization uses calendar revenue ÷ calendar spend for one closed day. Revenue that arrived yesterday over spend that happened yesterday, regardless of when those users installed.

Cohort ROAS (roas_d7, roas_d30) answers a different question — how a specific install cohort matures over time. Both are valid; they disagree sharply, and mixing them produces confident nonsense.

Teach the learner to ask: “am I judging a day, or judging a cohort?” The daily buy-side loop always judges a day.

Target is 120% per title, not to the portfolio average. A profitable portfolio hides losing titles inside it, and the portfolio number gives no action.

120% is not a margin target — it is the level at which a title is judged healthy enough to keep its current spend. A title at 114% is failing, even though it looks close.

Concrete: $126 spend returning $17 profit reads as “nearly at target” and is not healthy. The percentage flatters small numbers; see Lesson 3 for the fix.

Lesson 3 — Rank by dollar gap, not by ratio

Section titled “Lesson 3 — Rank by dollar gap, not by ratio”

Once titles are failing, order them by dollar gap, not by how far the percentage fell:

dollar gap = (spend × 1.20) − revenue

A title at 60% on $10 of spend is a worse ratio than a title at 110% on $2,000 — and fixing the second one is worth ~40× more money. Percentage ranking sends buyers to chase rounding error.

Work the funnel in this order, applying dollar-gap ranking at each level:

portfolio → titles (by dollar gap) → worst network inside the failing title → campaigns (by dollar gap)

Lesson 4 — Cuts size against actual spend

Section titled “Lesson 4 — Cuts size against actual spend”

Reduce budget against what actually got spent yesterday, never the nominal budget.

A campaign with a $60 budget that spent $39.65 is not constrained by $60. Cutting it to $50 changes nothing — delivery never reached the cap. The cut has to land below actual spend to bite; in that case, roughly $30.

Two more rules:

  • Budget reduction before pause. Pause is not the first move. Reduce, re-check tomorrow.
  • Magnitude is judgment, not a formula. Deliberately so — the right cut depends on how far off target, how noisy the title is, and how much history you have.

A title at or above 120% is an opportunity to spend more. Buyers trained only on cutting leave the portfolio smaller than it should be.

Two things generalize across networks:

  • The lever differs. Some networks scale by relaxing a ROAS goal (letting the algorithm buy more volume); others expose only whole-campaign budget and bid. Learn which lever the network actually gives you before deciding scaling is impossible.
  • Geos qualify on 7-day aggregate ROAS, not on yesterday and not on a count of winning days. One day is noise at geo volumes.

Campaign-level and geo-level verdicts are separate and often disagree — a campaign can hold at 132% yesterday, read 117% over 7 days, and still contain one geo worth scaling. Teach the learner that “the campaign is fine” and “this geo should scale” are not in conflict.

Braking a geo is the same judgement in reverse, with two traps. First, a geo that improved today is still failing if its 7-day aggregate is failing — measure against the target, not against yesterday’s version of itself. Second, know which direction your network’s lever runs: on some networks lowering a ROAS goal loosens the algorithm and buys more volume, so raising it is the brake. Confirm the direction before pulling it.

Rising spend into a failing geo jumps the queue. A flat-spend loser can wait a day. When spend is climbing day over day into something that has failed for a week, the algorithm is actively reallocating budget toward the loss and it compounds while you wait.

When the ad network’s dashboard and the MMP disagree on the same campaign-day, act on the network’s number.

Spend usually agrees to the cent; revenue does not, because each side counts what its own tracking sees. The network’s figure is both what the buyer sees and what the network’s bidder optimizes toward — so it is the number that predicts what the algorithm will do next.

A tool that reports the MMP’s (typically higher) figure will systematically recommend gentler action than the operator would take. That is a silent bias, not a rounding issue.

This is a decision rule, not a claim that one source is more truthful. Both are measuring honestly; they are measuring different things.

Training stops at understanding. To act, route to the skill that owns the mechanics:

Learner now wantsSkill
Run today’s triage loop on live datathe studio-local quick-optimize skill *
Diagnose a policy-limited Google Ads ad groupt1k-marketing-ua-how-to-optimize
Tune Mintegral bids, budgets, geos, sub-publisherst1k-marketing-ua-mintegral-campaign-optimization
Campaign structure and bidding strategy backgroundt1k-marketing-ua-campaigns

* quick-optimize is studio-local — not shipped with this kit. If it is not installed, the lessons still stand on their own; that handoff simply has no target and the learner acts from the fundamentals.

  • Do not let a training session turn into a live recommendation. The moment the learner supplies real numbers and asks “so what do I cut?”, hand off to the studio-local quick-optimize skill. Teaching material carries no freshness guarantee and this skill reads no data.
  • Intraday partial ROAS projects to close at +10 percentage points, flat — not ×1.10. The relative form inverts the conclusion on low numbers: 60% → 66% (still cut) versus 60% → 70% (still cut) looks similar, but 95% → 104.5% versus 95% → 105% changes the decision at the boundary. Teach the additive form and say why the multiplicative one is wrong.
  • Do not teach roas_dN as the daily metric. It is the single most common conversion error from buyers trained at cohort-reporting shops, and it survives a full explanation of Lesson 1 unless the learner works a case.
  • The 120% figure is a studio operating choice, not an industry constant. Teach it as “our current bar and the reasoning behind a bar”, so a learner does not carry it to a context with different margins.
  • Timing gates judgement. A title is not judgeable until the previous day’s data is fully landed (13:30 UTC+7 for this studio’s crawl). A learner who judges at 09:00 is reading a partial day and will cut healthy titles. Teach the timing before the rules.
  • Two handoff targets are local-only and will keep de-registering. quick-optimize and how-to-optimize exist on disk but have never been contributed to theonekit-marketing — they are registered only by a hand-edit to the local modules/ua/module.json, which every kit update replaces with the upstream copy. Each update silently drops them from keyword activation again. Invoke them by name if a keyword does not reach them, and treat contributing them upstream as the real fix. Anything registered only in a local manifest has this problem.